// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

$assert BATCH_TILE % 4 == 0
$assert BATCH_TILE >= 4
$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ"
$SSE_HEADER = {2: "emmintrin.h", 4: "smmintrin.h"}[SSE]
#include <assert.h>

#include <${SSE_HEADER}>

#include <xnnpack/vunary.h>
#include <xnnpack/common.h>


extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];

$ISA = {2: "sse2", 4: "sse41"}[SSE]
void xnn_f32_velu_ukernel__${ISA}_rr2_lut16_p3_x${BATCH_TILE}(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(n != 0);
  assert(n % sizeof(float) == 0);
  assert(x != NULL);
  assert(y != NULL);

  const __m128 vprescale = _mm_load_ps(params->sse2_rr2_lut16_p3.prescale);
  const __m128 valpha = _mm_load_ps(params->sse2_rr2_lut16_p3.alpha);
  const __m128 vbeta = _mm_load_ps(params->sse2_rr2_lut16_p3.beta);
  const __m128 vsat_cutoff = _mm_load_ps(params->sse2_rr2_lut16_p3.sat_cutoff);
  const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut16_p3.magic_bias);
  const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut16_p3.log2e);
  const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut16_p3.index_mask);
  const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_hi);
  const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_lo);
  const __m128 vc3 = _mm_load_ps(params->sse2_rr2_lut16_p3.c3);
  const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut16_p3.c2);
  const __m128 vone = _mm_load_ps(params->sse2_rr2_lut16_p3.one);

  $if BATCH_TILE > 4:
    for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) {
      __m128 vx${ABC[0:4]} = _mm_loadu_ps(x);
      $for N in range(4, BATCH_TILE, 4):
        __m128 vx${ABC[N:N+4]} = _mm_loadu_ps(x + ${N});
      x += ${BATCH_TILE};

      $for N in range(0, BATCH_TILE, 4):
        const __m128 vz${ABC[N:N+4]} = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx${ABC[N:N+4]}, vprescale));

      $for N in range(0, BATCH_TILE, 4):
        __m128 vn${ABC[N:N+4]} = _mm_add_ps(_mm_mul_ps(vz${ABC[N:N+4]}, vlog2e), vmagic_bias);

      $for N in range(0, BATCH_TILE, 4):
        const __m128i vidx${ABC[N:N+4]} = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn${ABC[N:N+4]}), vindex_mask), 2);
        const __m128i ven${ABC[N:N+4]} = _mm_slli_epi32(_mm_castps_si128(vn${ABC[N:N+4]}), 19);

      #if XNN_ARCH_X86_64
        $for N in range(0, BATCH_TILE, 4):
          const uint64_t vidx${ABC[N:N+2]} = (uint64_t) _mm_cvtsi128_si64(vidx${ABC[N:N+4]});
          $if SSE >= 4:
            const uint64_t vidx${ABC[N+2:N+4]} = (uint64_t) _mm_extract_epi64(vidx${ABC[N:N+4]}, 1);
          $else:
            const uint64_t vidx${ABC[N+2:N+4]} = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx${ABC[N:N+4]}, vidx${ABC[N:N+4]}));
          const __m128i vl${ABC[N]}   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N:N+2]})));
          const __m128i vl${ABC[N+2]} = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx${ABC[N+2:N+4]})));
          $if SSE >= 4:
            const __m128i vl${ABC[N:N+2]} = _mm_insert_epi32(vl${ABC[N]}, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N:N+2]} >> 32))), 1);
          $else:
            const __m128i vl${ABC[N+1]} = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N:N+2]} >> 32))));
            const __m128i vl${ABC[N:N+2]} = _mm_unpacklo_epi32(vl${ABC[N]}, vl${ABC[N+1]});
          $if SSE >= 4:
            const __m128i vl${ABC[N+2:N+4]} = _mm_insert_epi32(vl${ABC[N+2]}, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N+2:N+4]} >> 32))), 1);
          $else:
            const __m128i vl${ABC[N+3]} = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx${ABC[N+2:N+4]} >> 32))));
            const __m128i vl${ABC[N+2:N+4]} = _mm_unpacklo_epi32(vl${ABC[N+2]}, vl${ABC[N+3]});
          const __m128i vl${ABC[N:N+4]} = _mm_unpacklo_epi64(vl${ABC[N:N+2]}, vl${ABC[N+2:N+4]});
      #else  // !XNN_ARCH_X86_64
        $for N in range(0, BATCH_TILE, 4):
          const uint32_t vidx${ABC[N]} = (uint32_t) _mm_cvtsi128_si32(vidx${ABC[N:N+4]});
          const uint32_t vidx${ABC[N+1]} = (uint32_t) _mm_extract_epi16(vidx${ABC[N:N+4]}, 2);
          const uint32_t vidx${ABC[N+2]} = (uint32_t) _mm_extract_epi16(vidx${ABC[N:N+4]}, 4);
          const uint32_t vidx${ABC[N+3]} = (uint32_t) _mm_extract_epi16(vidx${ABC[N:N+4]}, 6);
          const __m128i vl${ABC[N]}   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx${ABC[N]})));
          const __m128i vl${ABC[N+2]} = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx${ABC[N+2]})));
          $if SSE >= 4:
            const __m128i vl${ABC[N:N+2]} = _mm_insert_epi32(vl${ABC[N]}, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx${ABC[N+1]})), 1);
          $else:
            const __m128i vl${ABC[N+1]} = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx${ABC[N+1]})));
            const __m128i vl${ABC[N:N+2]} = _mm_unpacklo_epi32(vl${ABC[N]}, vl${ABC[N+1]});
          $if SSE >= 4:
            const __m128i vl${ABC[N+2:N+4]} = _mm_insert_epi32(vl${ABC[N+2]}, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx${ABC[N+3]})), 1);
          $else:
            const __m128i vl${ABC[N+3]} = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx${ABC[N+3]})));
            const __m128i vl${ABC[N+2:N+4]} = _mm_unpacklo_epi32(vl${ABC[N+2]}, vl${ABC[N+3]});
          const __m128i vl${ABC[N:N+4]} = _mm_unpacklo_epi64(vl${ABC[N:N+2]}, vl${ABC[N+2:N+4]});
      #endif  // XNN_ARCH_X86_64

      $for N in range(0, BATCH_TILE, 4):
        vn${ABC[N:N+4]} = _mm_sub_ps(vn${ABC[N:N+4]}, vmagic_bias);
        __m128 vs${ABC[N:N+4]} = _mm_castsi128_ps(_mm_add_epi32(vl${ABC[N:N+4]}, ven${ABC[N:N+4]}));

      $for N in range(0, BATCH_TILE, 4):
        __m128 vt${ABC[N:N+4]} = _mm_add_ps(_mm_mul_ps(vn${ABC[N:N+4]}, vminus_ln2_hi), vz${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vt${ABC[N:N+4]} = _mm_add_ps(_mm_mul_ps(vn${ABC[N:N+4]}, vminus_ln2_lo), vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        __m128 vp${ABC[N:N+4]} = _mm_add_ps(_mm_mul_ps(vc3, vt${ABC[N:N+4]}), vc2);

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = _mm_mul_ps(vp${ABC[N:N+4]}, vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        vt${ABC[N:N+4]} = _mm_mul_ps(vt${ABC[N:N+4]}, vs${ABC[N:N+4]});
        vs${ABC[N:N+4]} = _mm_sub_ps(vs${ABC[N:N+4]}, vone);

      $for N in range(0, BATCH_TILE, 4):
        vp${ABC[N:N+4]} = _mm_add_ps(_mm_mul_ps(vp${ABC[N:N+4]}, vt${ABC[N:N+4]}), vt${ABC[N:N+4]});

      $for N in range(0, BATCH_TILE, 4):
        const __m128 ve${ABC[N:N+4]} = _mm_mul_ps(_mm_add_ps(vp${ABC[N:N+4]}, vs${ABC[N:N+4]}), valpha);

      $for N in range(0, BATCH_TILE, 4):
        $if SSE < 4:
          const __m128 vm${ABC[N:N+4]} = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx${ABC[N:N+4]})));
        vx${ABC[N:N+4]} = _mm_mul_ps(vx${ABC[N:N+4]}, vbeta);

      $for N in range(0, BATCH_TILE, 4):
        $if SSE >= 4:
          const __m128 vy${ABC[N:N+4]} = _mm_blendv_ps(vx${ABC[N:N+4]}, ve${ABC[N:N+4]}, vx${ABC[N:N+4]});
        $else:
          const __m128 vy${ABC[N:N+4]} = _mm_or_ps(_mm_and_ps(ve${ABC[N:N+4]}, vm${ABC[N:N+4]}), _mm_andnot_ps(vm${ABC[N:N+4]}, vx${ABC[N:N+4]}));

      _mm_storeu_ps(y, vy${ABC[0:4]});
      $for N in range(4, BATCH_TILE, 4):
        _mm_storeu_ps(y + ${N}, vy${ABC[N:N+4]});
      y += ${BATCH_TILE};
    }
  for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
    __m128 vx = _mm_loadu_ps(x);
    x += 4;

    const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));

    __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);

    const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
    const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
    #if XNN_ARCH_X86_64
      const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
      $if SSE >= 4:
        const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
      $else:
        const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
      const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
      const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
      $if SSE >= 4:
        const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))), 1);
      $else:
        const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
        const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
      $if SSE >= 4:
        const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))), 1);
      $else:
        const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
        const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
    #else  // !XNN_ARCH_X86_64
      const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
      const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
      $if SSE >= 4:
        const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
      $else:
        const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
        const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
      $if SSE >= 4:
        const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
      $else:
        const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
        const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
    #endif  // XNN_ARCH_X86_64
    const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
    __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
    vn = _mm_sub_ps(vn, vmagic_bias);

    __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
    vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);

    __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
    vp = _mm_mul_ps(vp, vt);

    vt = _mm_mul_ps(vt, vs);
    vs = _mm_sub_ps(vs, vone);
    vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
    const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);

    $if SSE < 4:
      const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
    vx = _mm_mul_ps(vx, vbeta);
    $if SSE >= 4:
      const __m128 vy = _mm_blendv_ps(vx, ve, vx);
    $else:
      const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));

    _mm_storeu_ps(y, vy);
    y += 4;
  }
  if XNN_UNLIKELY(n != 0) {
    __m128 vx = _mm_loadu_ps(x);

    const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));

    __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);

    const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
    const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
    #if XNN_ARCH_X86_64
      const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
      $if SSE >= 4:
        const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
      $else:
        const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
      const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
      const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
      $if SSE >= 4:
        const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))), 1);
      $else:
        const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
        const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
      $if SSE >= 4:
        const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))), 1);
      $else:
        const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
        const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
    #else  // !XNN_ARCH_X86_64
      const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
      const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
      $if SSE >= 4:
        const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
      $else:
        const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
        const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
      $if SSE >= 4:
        const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
      $else:
        const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
        const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
    #endif  // XNN_ARCH_X86_64
    const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
    __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
    vn = _mm_sub_ps(vn, vmagic_bias);

    __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
    vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);

    __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
    vp = _mm_mul_ps(vp, vt);

    vt = _mm_mul_ps(vt, vs);
    vs = _mm_sub_ps(vs, vone);
    vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
    const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);

    $if SSE < 4:
      const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
    vx = _mm_mul_ps(vx, vbeta);
    $if SSE >= 4:
      __m128 vy = _mm_blendv_ps(vx, ve, vx);
    $else:
      __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));

    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vy);
      vy = _mm_movehl_ps(vy, vy);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vy);
    }
  }
}